Data-Driven Analytics for Manufacturing and Healthcare

Nowadays, healthcare and manufacturing are becoming increasingly important in our lives. As human life expectancy increases and the population ages, age-related chronic diseases are increasing, and the associated medical expenses are also increasing year by year. Therefore, there is an urgent need for developing novel cheap and smarter strategies of providing health care to people suffering from such chronic diseases. Similarly, the manufacturing process is mainly divided into two types, traditional mass production and personalized production. With the increasing population, the emergence and development of e-shopping has expanded the market and the demand for products has also increased. Industries have shifted to manufacturing methods based on lean and customers' demand. Therefore, it becomes more challenging to adopt systematic methodologies for monitoring and preventing defects occurrence during the manufacturing process. Furthermore, due to customization, the time for the production line to be optimized has been significantly reduced lead to an increase in defective products. Hence, more sophisticated quality management strategies for manufacturing are needed in order to meet customers' demand. To fill in these gaps, this dissertation has three research topics. The first topic is about proposing a novel sequential decision-making framework to derive the optimal cancer intervention strategies for BRCA mutation carriers. Markov Decision Process (MDP) to model the dynamic progression of cancer risk for the mutation carriers has been solved using Monte Carlo Tree Search (MCTS). Experimental results demonstrate that the proposed MCTS-planning method can effectively provide optimal intervention strategies for BRCA mutation carriers which reduces the cancer incidence risk and maintains a high-level QOL at the same time. The second topic is about detecting the surface defection of aluminum die-casting parts. Traditional Convolutional Neural Network (CNN), pre-trained ResNet-18, and pre-trained Vision Transformer (ViT) models are utilized to detect defective parts from produced products. We evaluate the model performance through four metrics accuracy, precision, recall, and F1-score. Model performances demonstrate that the implemented three methods can effectively detect defective parts from the produced aluminum lids. The third topic is quality control for 3D-printed parts. We propose a four-phase framework to determine the robust range and level of sensitive FDM process parameters for achieving the desired tensile strength. We implement and validate the first three phases, demonstrating the framework's effectiveness. Experimental results demonstrate that the proposed framework effectively identifies sensitive parameters and determines the appropriate range or level values for achieving optimal or acceptable tensile strength.